Audio-to-Notes Transcriber — Melody & Sheet Music

One instrument at a time, played cleanly, with the notes clearly separated — a solo piano, a picked guitar, a sung melody without accompaniment. The network was built to find where each note begins and what pitch it is, and everything that blurs those two things costs accuracy: reverb smears the boundary between one note and the next, a full band puts four instruments in the same frequency range at once, and heavy distortion adds overtones that read as extra notes. Two instruments playing together are the honest limit — the result will contain both, interleaved, with no way to tell which line belonged to which. Separate the parts first if you have them; if you do not, expect a sketch rather than a score.

We measured it against reference recordings whose notes are known, rather than estimating it. One instrument or voice: 65-80% of the notes come out right. A dense mix of several instruments: 10-40%, and the melody line we highlight there is a sketch rather than the tune — for songs the Chords tab is more dependable than the noteheads.

MIDI is a universal format for musical notes. It can be opened in any DAW (GarageBand, Ableton, FL Studio) or notation software (MuseScore, Sibelius).

Yes. If the song has singing, a separate neural network finds the vocal melody straight from the full recording: on most computers this happens by itself after the transcription, otherwise with the “Refine the melody” button. Check it by ear: fast runs can come out simplified. In an instrumental song the melody will be approximate — some notes may come from the accompaniment. All detected notes are under “All notes”, the harmony is on the “Chords” tab; a single instrument or voice without music is transcribed most accurately.

Yes. Paste the song’s lyrics into the “Chords over lyrics” block under the result — one line per line of the song — and check “Show words on the stave”: the syllables land under the melody notes, the chords above the staff, and the sheet exports that way to PDF and MusicXML. This tool does not recognize the words from the singing by itself; if you don’t have them, you can get them with “Lyrics Transcription from Audio”.

A MIDI file and a piano roll, plus generated notation you can look at. MIDI is the useful one: it holds the notes as events rather than as an image, so anything musical will take it — a notation editor for cleaning up and printing, a DAW for putting the part on a different instrument, a phone app for slowing it down. The notation is a rendering of the same data and is best treated as a reading aid rather than as a finished score, because turning a performance into readable rhythm involves decisions that a machine makes badly: a note played slightly late becomes a rest and a dotted figure. Import the MIDI into a notation program and quantise it there — that is where those decisions belong to you.

Yes. Open the Chords tab to see the chord progression over time, jump to any chord, and copy or download it. It works best on clear single-instrument recordings.

Yes. After the transcription, press «For piano»: the right hand gets the melody, the left hand gets the bass and chords from the detected harmony, on a grand staff. Neither hand reaches wider than an octave, but the notes are approximate — check them by ear; details of the original arrangement are simplified. In a backing track without vocals, the lead instrument becomes the melody. Download the part as PDF, MIDI or MusicXML.

Yes. After the transcription, answer “Choir or ensemble” or press “By part”: all detected notes are split into parts from top to bottom, each part gets its own staff, and you can change the number of parts. On a computer with a graphics card, a choir network then refines the parts: on open choir recordings with scores, it puts more than 90 % of each part where it belongs with two voices and about 85 % with four; without it, about 80 % and 60 %. Where parts sing the same note or cross, they may swap places, so check by ear. The score downloads as PDF, MusicXML (one part per voice) and MIDI (one track per voice).

The whole recording is transcribed, and without Premium its first half is unlocked; a short recording is unlocked in full. The rest is shown locked on the piano roll: Premium unlocks it right away, with no second transcription, together with the chords, the piano arrangement and export of the whole recording.